Online learning is machine learning, in real time from successive data samples. Meta online learning consists in combining several online learning algorithms from a given set (termed portfolio) of algorithms. The goal can be (i) mitigating the effect of a bad choice of online learning algorithms (ii) parallelization (iii) combining the strengths of different algorithms. Basically, meta online learning boils down to combining noisy optimization algorithms. Whereas many tools exist for combining combinatorial optimization tools, little is known about combining noisy optimization algorithms. Recently, a methodology termed lag has been proposed for that. We test experimentally the lag methodology for online learning, for a stock management problem and a cartpole problem.